AI Models Still Fail to Properly Correct Users, Study Finds

A joint study by UW, AI2, LMU Munich, UC Berkeley, Stanford and UW-Madison found that even when AI models recognize a user's false claim, they often fail to correct it clearly, burying objections in praise and partial agreement. In 29.4% of information-seeking responses models mixed correction with affirmation, rising to 60.1% for Gemini 2.5 Pro.

• 29.4% of responses mixed correction with affirmation; 60.1% for Gemini 2.5 Pro
• Responses averaged 46.7 lines, with corrective phrases in only 2–3
• Seven models and five languages tested: English, German, Chinese, Hindi, Yoruba
• A 'wait a minute' intervention increased pushback but also false objections